NVIDIA Open-Sources Massive Multimodal Training Data: 10 Trillion Language Tokens, 500K Robot Trajectories, 455K Protein Structures

NVIDIA Blog·medium signal

Alongside the Computex model announcements, NVIDIA released one of the world's largest open multimodal data collections: 10 trillion language training tokens, 500,000 robotics trajectories, 455,000 protein structures, and 100 terabytes of vehicle sensor data. The dataset spans the Nemotron, Cosmos, Isaac GR00T, Clara, and Alpamayo model families. For builders, the robotics trajectories and protein structures represent previously gated data categories now accessible under open licenses.